Machine-Learning Interatomic Potential Manager

Compute descriptors

Descriptor

class mlacs.mlip.Descriptor(atoms, rcut=5.0, alpha=1.0, prefix='MLIP', **kwargs)

Base class for descriptors

Parameters:

  • atoms (ase.atoms) – Reference structure, with the elements for the descriptor

  • rcut (float) – The cutoff for the descriptor

SumDescriptor

class mlacs.mlip.SumDescriptor(*args)

A class to mix several descriptors together.

Parameters:

args (list of Descriptor) – A list of all the descriptors to mix.

SnapDescriptor

class mlacs.mlip.SnapDescriptor(atoms, rcut=5.0, parameters={}, model='linear', alpha=1.0, alpha_quad=1.0, prefix='SNAP', **kwargs)

Interface to the SNAP potential of LAMMPS.

Parameters:

  • atoms (ase.atoms) – Reference structure, with the elements for the descriptor

  • rcut (float) – The cutoff of the descriptor, in angstrom Default 5.0

  • parameters (dict) –

    A dictionnary of parameters for the descriptor input

    The default values are

    • twojmax = 8

    • rfac0 = 0.99363

    • rmin0 = 0.0

    • switchflag = 1

    • bzeroflag = 1

    • wselfallflag = 0

  • model (str) – The type of model use. Can be either ‘linear’ or ‘quadratic’ Default linear

  • alpha (float) – The multiplication factor to the regularization parameter for ridge regression. Default 1.0

  • alpha_quad (float) – A multiplication factor for the regularization that apply only to the quadratic component of the descriptor Default 1.0

Examples

>>> from ase.build import bulk
>>> atoms = bulk("Cu", cubic=True).repeat(2)
>>>
>>> from mlacs.mlip import SnapDescriptor, LinearPotential
>>> desc = SnapDescriptor(atoms, rcut=4.2, parameters=dict(twojmax=6))
>>> desc.compute_descriptor(atoms)

MliapDescriptor

class mlacs.mlip.MliapDescriptor(atoms, rcut=5.0, parameters={}, model='linear', style='snap', alpha=1.0, prefix='MLIAP', **kwargs)

Interface to the MLIAP potential of LAMMPS.

Parameters:

  • atoms (ase.atoms) – Reference structure, with the elements for the descriptor

  • rcut (float) – The cutoff of the descriptor, in angstrom Default 5.0

  • parameters (dict) –

    A dictionnary of parameters for the descriptor input

    If the style is set to snap, then the default values are

    • twojmax = 8

    • rfac0 = 0.99363

    • rmin0 = 0.0

    • switchflag = 1

    • bzeroflag = 1

    • wselfallflag = 0

    If the style is set to so3, then the default values are

    • nmax = 4

    • lmax = 4

    • alpha = 1.0

  • model (str) – The type of model use. Can be either ‘linear’ or ‘quadratic’ Default linear

  • style (str) – The style of the descriptor used. Can be either ‘snap’ or ‘so3’ Default ‘snap’

  • alpha (float) – The multiplication factor to the regularization parameter for ridge regression. Default 1.0

Examples

AceDescriptor

class mlacs.mlip.AceDescriptor(atoms, free_at_e, rcut=5.0, tol_e=5, tol_f=25, bconf_dict=None, loss_dict=None, fitting_dict=None, backend_dict=None, nworkers=None)

Interface to the ACE potential using python-ace. Pace potential in LAMMPS. Usually a better generalization than SNAP but harder to train. Can be used with state at different pressure. Note : You will probably want to use 10-20 states since the

fitting is relatively slow

Parameters:

  • atoms (ase.atoms or list) – Reference structure, with the elements for the descriptor

  • free_at_e (dict) – The energy of one atom for every element. The atom MUST be isolated in a big box (eV/at) e.g. : {‘Cu’: 12.1283, ‘O’: 3.237}

  • rcut (float) – The cutoff of the descriptor, in angstrom Default 5.0

  • tol_e (float) – The tolerance on energy between ACE and DFT (meV/at) Default 5

  • tol_f (float) – The tolerance on forces between ACE and DFT (meV/ang) Default 25

  • bconf_dict (dict) –

    A dictionnary of parameters for the BBasisConfiguration The default values are

    • deltaSplineBins: 0.001

    • elements: Found dynamically (e.g. Cu2O: [‘Cu’, ‘O’])

    • embeddings: {“ALL”: {npot: ‘FinnisSinclairShiftedScaled’,

      fs_parameters: [1, 1, 1, 0.5], ndensity: 2}}

    • bonds: {“ALL”: {radbase: ‘ChebExpCos’,

      NameOfCutoffFunction: ‘cos’, radparameters: [5.25], rcut: 7, dcut: 0.01}}

    • functions: {“ALL”: {nradmax_by_orders: [15, 3, 2],

      lmax_by_orders: [0, 2, 2]}}

  • loss_dict (dict) –

    A dictionnary of parameters for the loss function The default values are

    • kappa: 0.3

    • L1_coeffs: 1e-12

    • L2_coeffs: 1e-12

    • w0_rad: 1e-12

    • w1_rad: 1e-12

    • w2_rad: 1e-12

  • fitting_dict (dict) –

    A dictionnary of parameters for the minimization The default values are

    • weighting: MBAR|Uniform otherwise

    • loss: loss_dict

    • maxiter: 100

    • fit_cycles: 1

    • repulsion: ‘auto’

    • optimizer: ‘BFGS’

    • optimizer_options: {disp: True, gtol: 0, xrtol: 0}

  • backend_dict (dict) –

    A dictionnary of parameters for the backend The default values are

    • evaluator: ‘tensorpot’

    • parallel_mode: ‘parallel’

    • n_workers : None

    • batch_size: 100

    • display_step: 50

Add a weighting policy

FixedWeight

class mlacs.mlip.FixedWeight(static_weight=array([0]), subweight=None, energy_coefficient=1.0, forces_coefficient=1.0, stress_coefficient=1.0)

Class that gives a static weight to the first few configurations and then use the given WeightingPolicy for the other configurations. Can be used to give a fixed weight to the training configurations

Parameters:

  • subweight (WeightingPolicy) – Weight of the configurations after Default UniformWeight

  • static_weight (np.array) – Weights of the first configuration. The number of configuration is given by the length of the array. The weight must be normalized Default [0]

EnergyBasedWeight

class mlacs.mlip.EnergyBasedWeight(delta=1.0, energy_coefficient=1.0, forces_coefficient=1.0, stress_coefficient=1.0, database=None, weight=None)

Class that gives weight according to w_n = C/[E_n - E_min + delta]**2 where C is a normalization constant.

Parameters:

delta (float) – Shift to avoid overweighting of the ground state (eV/at) Default 1.0

UniformWeight

class mlacs.mlip.UniformWeight(nthrow=0, energy_coefficient=1.0, forces_coefficient=1.0, stress_coefficient=1.0, database=None, weight=None)

Class that gives uniform weight in MLACS.

Parameters:

nthrow (int) –

Number of configurations to ignore when doing the fit. Three cases:

  1. If nconf > 2*nthrow, remove the nthrow first configuration

  2. If nthrow < nconf < 2*nthrow, remove the nconf-nthrow first conf

  3. If nconf < nthrow, keep all conf

IncreasingWeight

class mlacs.mlip.IncreasingWeight(nthrow=0, power=1, energy_coefficient=1.0, forces_coefficient=1.0, stress_coefficient=1.0, database=None, weight=None)

Class that gives increasing weight with the index of a configuration in MLACS. This weighting policy has been though for structural optimization.

Parameters:

nthrow (int) –

Number of configurations to ignore when doing the fit. Three cases :

  1. If nconf > 2*nthrow, remove the nthrow first configuration

  2. If nthrow < nconf < 2*nthrow, remove the nconf-nthrow first conf

  3. If nconf < nthrow, keep all conf

MbarManager

class mlacs.mlip.MbarManager(parameters={}, energy_coefficient=1.0, forces_coefficient=1.0, stress_coefficient=1.0, **kwargs)

Computation of weight according to the multistate Bennett acceptance ratio (MBAR) method for the analysis of equilibrium samples from multiple arbitrary thermodynamic states.

Parameters:

  • solver (str) – Define type of solver for pymbar Default L-BFGS-B

  • scale (float) – Imposes weights for the new configurations. Default 1.0

  • start (int) – Step to start weight computation. At least 2 since you need two potentials to compare them. Default 2

  • database (ase.Trajectory) – Initial database (optional) Default None

  • weight (list or str) – If you use an initial database, it needs weight. Can a list or an np.array of values or a file. Default None

Fitting model

MlipManager

class mlacs.mlip.MlipManager(descriptor, weight=None, folder='MLIP', **kwargs)

Parent Class for the management of Machine-Learning Interatomic Potential

LinearPotential

class mlacs.mlip.LinearPotential(descriptor, parameters={}, weight=None, **kwargs)

Potential that assume a linear relation between the descriptor and the energy.

Parameters:

  • descriptor (Descriptor) – The descriptor used in the model.

  • parameters (dict) – The parameters for the fit. By default, the fit is a simple ordinary least squares. Ridge regression can be use by setting a dictionnary as dict(method=ridge, lambda_ridge=alpha), with alpha the ridge coefficient.

  • weight (WeightingPolicy) – Weight used for the fitting and calculation of properties. Default None, which results in the use of uniform weights.

Examples

>>> from ase.io import read
>>> confs = read('Trajectory.traj', index=':')
>>>
>>> from mlacs.mlip import SnapDescriptor, LinearPotential
>>> desc = SnapDescriptor(confs[0], rcut=6.2, parameters=dict(twojmax=6))
>>> mlip = LinearPotential(desc)
>>>
>>> mlip.update_matrices(confs)
>>> mlip.train_mlip()

MomentTensorPotential

class mlacs.mlip.MomentTensorPotential(atoms, mlpbin='mlp', mtp_parameters={}, fit_parameters={}, folder='MTP', **kwargs)

Interface to the Moment Tensor Potential of the MLIP package.

Parameters:

  • atoms (ase.Atoms) – Prototypical configuration for the MLIP. Should have the desired species.

  • mlpbin (str) – The path to the mlp binary. If mpi is desired, the command should be set as ‘mpirun /path/to/mlp’

  • mtp_parameters (dict) –

    The dictionnary with inputs for the potential.

    The default values are set to

    • level = 8

    • radial_basis_type = ‘RBChebyshev’

    • min_dist=1.0,

    • max_dist=5.0,

    • radial_basis_size=8

  • fit_parameters (dict) –

    The parameters for the fit of the potential

    The default parameters are set to

    • scale_by_forces=0

    • max_iter=1000

    • bfgs_conv_tol=1e-3

    • weighting=’vibrations’

    • (weight_scaling=1, weight_scaling_forces=0 for MLIP-3)

    • init_params=’random’

    • update_mindist=False

Examples

>>> from ase.io import read
>>> confs = read('Trajectory.traj', index=':')
>>>
>>> from mlacs.mlip import MomentTensorPotential
>>> mlip = MomentTensorPotential(confs[0], mtp_parameters=dict(level=6))
>>> mlip.update_matrices(confs)
>>> mlip.train_mlip()

TensorpotPotential

class mlacs.mlip.TensorpotPotential(descriptor, weight=None, **kwargs)

Potential that use Tensorpotential to minimize a cost function. For now, only works with AceDescriptor

Parameters:

  • descriptor (Descriptor) – The descriptor used in the model.

  • weight (WeightingPolicy) – Weight used for the fitting and calculation of properties. Default None

Examples

>>> from ase.io import read
>>> confs = read('Trajectory.traj', index=':')
>>>
>>> from mlacs.mlip import AceDescriptor, TensorpotPotential
>>> desc = AceDescriptor(confs[0], free_at_e={'Cu': 0}, rcut=rcut)
>>> mlip = TensorpotPotential(desc, folder="ACE")
>>>
>>> mlip.update_matrices(confs)
>>> mlip.train_mlip()

DeltaLearningPotential

class mlacs.mlip.DeltaLearningPotential(model, pair_style, pair_coeff, model_post=None, atom_style='atomic', folder=None, **kwargs)
Parameters:

  • model (MlipManager) – The MLIP model to train on the difference between the true energy and the energy of a LAMMPS reference model.

  • pair_style (str or list of str) – The pair_style of the LAMMPS reference potential. If only one pair style is used, can be set as a str. If an overlay of pair style is used, this input as to be a list of str of the pair_style. For example :

  • pair_coeff (list of str) – The pair_coeff of the LAMMPS reference potential.

  • folder (str (optional)) – The root for the directory in which the MLIP are stored. Default ‘MLIP’

Examples

>>> from ase.io import read
>>> confs = read('Trajectory.traj', index=':')
>>>
>>> from mlacs.mlip import SnapDescriptor, LinearPotential
>>> desc = SnapDescriptor(confs[0], rcut=6.2, parameters=dict(twojmax=6))
>>> mlip = LinearPotential(desc)
>>>
>>> from mlacs.mlip import DeltaLearningPotential
>>> ps = ['sw', 'zbl 3.0 4.0']
>>> pc = [['* * Si.sw Si'], ['* * 14 14']]
>>> dmlip = DeltaLearningPotential(mlip, pair_style=ps, pair_coeff=pc)
>>> dmlip.update_matrices(confs)
>>> dmlip.train_mlip()

MlipCalculator

class mlacs.mlip.calculator.MlipCalculator(model, **kwargs)

Ase Caculator object for MLACS generated MLIP that can take atomic configurations from an Atoms object and calculate the energy, forces and also stresses.